`missing_dims` option for aggregation methods like `mean` and `std`

[jbusecke]

I work a lot with climate model output and often loop over several models, of which some have a 'member' dimension and others don't.

I end up writing many lines like this:

for ds in model_datasets:
    if 'member_id' in ds.dims:
        ds = ds.mean('member_id)

Which often makes for very lengthy code blocks.

I recently noticed that .isel() actually has a nifty keyword argument 'missing_dims', which enables the user to apply isel and it just doesn't do anything when the dimension is not present.

I'd love to be able to do:

for ds in model_datasets:
    ds = ds.mean('member_id', missing_dims='ignore')

Is there a way to implement this generally for xarray aggregation methods (mean/max/min/std/...). Or is there a reason this should be avoided?

[max-sixty]

That seems like a reasonable suggestion @jbusecke .

To confirm, would ds.groupby('lat', 'long').mean(...) work? i.e. are the dimensions you don't want to reduce over reliable?

[dcherian]

Alternatively, you could run the following at the beginning

# not sure if syntax is right
model_datasets = [
    ds.expand_dims('member_id') 
    if "member_id" not in ds.coords else ds 
    for ds in model_datasets
]

so all your datasets are consistent.

[TomNicholas]

I ran into the same sort of thing today, when trying to loop over many datasets (each of which contained the contents of a node in a datatree...).

I also think that adding a missing_dims argument to all the array reduce methods would be useful, and I plan to have a go at it.

[dcherian]

My concern is that we could conceivably adding missing_dims to any function that takes a dim argument, which is pretty much the whole API.

For datatree, you could apply the reduction with the set-intersection of provided dims and dims present in a node (if that's the right term).

[TomNicholas]

For datatree, you could apply the reduction with the set-intersection of provided dims and dims present in a node (if that's the right term).

I specifically want the user to be able to choose between different behaviours with a flag, but you're right that I could just deal with this at the datatree level instead of here. That would make a fair amount of sense, and it would cover Julius' use-case (via encouraging him to store his models in a tree, so that for ds in model_datasets would become a loop over nodes in a tree).

My concern is that we could conceivably adding missing_dims to any function that takes a dim argument, which is pretty much the whole API.

Do you think that's a problem though? We added keep_attrs to even more of the API than this would cover. Specifically I would want to add it to the REDUCE_METHODS, the NAN_REDUCE_METHODS, and the NAN_CUM_METHODS (so {"all", "any", "max", "min", "mean", "prod", "sum", "std", "var", "median", "cumsum", "cumprod"}).

I'm fine with doing it either here or in datatree personally.